An analytical comparison of nearest neighbor algorithms for load balancing in parallel computers. Creative Commons: Attribution 3.0 Hong Kong License
|
|
- Marcia Melton
- 8 years ago
- Views:
Transcription
1 Title An analytical comparison of nearest neighbor algorithms for load balancing in parallel computers Author(s) Xu, CZ; Monien, B; Luling, R; Lau, FCM Citation The 9th International Parallel Processing Symposium Proceedings, Santa Barbara, CA, USA, April 1995, p Issued Date 1995 URL Rights Creative Commons: Attribution 3.0 Hong Kong License
2 An analytical comparison of nearest neighbor algorithms for load balancing in parallel computers Chengzhong Xu Department of Eiectrccal and Computer Engg. Wayne State Unavrrsrty, MI Burkhard Monien, Reinhard Luling Department of Computer Sccence Unrversrty of Paderborn, Germany Francis C. M. Lau Department oj Computer Scrence, The University of Hong h'ong, Hong Kong Abstract With nearest nezghbor load balanczng algorzthms, a processor makes balancang decasaons based on ats local anformatron and manages workload magrataons wathan at5 netghborhood. Thzs paper compares a couple of faarly well-known nearest neayhbor algorithms, the dimension exchange and the diffusion methods and thew varaants an terms of thew performances an both one-port and all-port communzcatzon archztdures. It turns out thirt the damensaon ~xchairye method outperforms the daffusion method an the ont -port continunacataon nwdei, and that the strength of the daffuszon method zs an asynchronous amplementataons an the all-port communicataon model The underlying comrnunrcataon networks coiisadered assume the most popular topologzes, the mesh a d the torus and thew bperad cases the hypercubu and the 6-ary n-cube. 1 Introduction Massively parallel computers have been shown to be very efficient at solving problems that can be partitioned into tasks with static compukation and cominunication patterns. However, there exist a large class of problems that have unpredictable computational requirements and/or irregular commiinication patterns. To solve this kind of problems efficiently in parallel comput ers, it is necessary to perform load balancing operatioiis during run time. Nearest neighbor load balancing algorithms are a cla.5s of methods in which processors make decisions based on local information in a decentralized manner and manage workload migrations within their neighborhood [l, 21. Since they have a less stringent requirement on the spread of local workload around the system, they are scalable to support massively parallel computers and suitable for retaining the communication locality inherent in the underlying computations. They are also iterative in nature in the sense that successively imposing local load balancing makes progress t,owards a global balanced state, and hence flexible in controlling the balance quality over the spectrum from the objective of load sharing that assures no idle processors coexist together with busy processors tl, the degree of global balanced state. Nearest neighbor load balancing algorithms rely on successive approximation to a global uniform distribution, and hence at each operation, need only be concerned with the direction of workload migration and the issue of how to apportion excess workloads. There a.re a number of ways for the choice of the direction of workload migration. Among of them, we are inter-. ested in a couple of simple representatives, fhe diflusion (DF, for short) and the dimension exchange (DE for short) methods. With the diffusion method, a highly or lightly loaded processor balanc.es its workload with all of its nearest neighhors simultaneously in a load balancing operation [3, 41. With the the dimension exchange method, by contrast, a processor in need of load balancing balances its workload successively with its neighbors one at a time and it,s new workload index will be considered in the the subsequent pairwise balancing [3, 5, 61. 'They are closely related because they lend themselves particularly well to implementation in two basic communication architectures, the all-port and the one-port models, respectively. The all-port model allows a processor to exchange messages with all its direct neighbors simultaneously in a communication step, while the oneport model restricts a processor t80 exchange messages with at most one direct neighbor at a time. Both of these two models are valid in real parallel computers / IEEE: 472
3 and were assumed in many recent, researches on communication algorithms ([7], for example). The all-port and one-port models favor the diffusion and the dimension exchange methods, respectively. In a system that supports d-port concurrent communications, a load balancing operation using the diffusion method can be completed in one communication step while that using the dimension exchange method would take d steps. It appears that the diffusion method has an advantage over the dimension method as far as exploiting the communication bandwidth is concerned. A natural but interesting question is whether the advantage translates into real performance benefits in load balancing or not. The performance of a load balancing algorithm is determined by t8wo measures. One is elgiciency which is reflected by the number of communication steps required by the algorithm to drive an initial workload distrisution into a uniform distribution. This measure alone is sufficient for those kinds of problems that need global balancing at run time. However, for the other kinds of applications that need to achieve load sharing rather than global balancing, we need another measure, the balance qualaty, to reflect the ability of the algorithm in bounding the variance of processors' workloads after performing one or more load balancing operations. The objective of this study is to answer the question concerning tlie performance of thp diffusion and the dimension exchange methods in different communication models. In this paper, we make a comprehensive comparison between the diffusion and the dimension exchange methods with respect to their efficiencies and balancing qualities when they are implemented in both oneport and all-port communication models, using synchronous/asynchronous invocation policies, and with static/dynamic random workload behaviors The communication networks to be considered include the structures of n-d torus and mesh, and their special cases: the ring, the chain, the hypercube and the IC-ary n-cube. We limit our scope to these structures because they are the most popular choices of topologies in commercial parallel computers [SI. Both the dimension exchange and the diffusion methods are parameterized algorithms, and their performance is largely influenced by the choice of the parameter values. We focus on two choices of the parameter value in each method: the iwerage DE (ADE), the optimally-tuned DE (ODE), the local average DF (ADF), and the optimally-tuned DF (ODF). The optimality here is in terms of the efficiency in static synchronous implementations among various choices of the DE and the DF parameters The average versions (ADE and ADF) are the most original versions and are still being employed in real applications today; we therefore include them in our comparison. Our main results are that the dimension exchange method outperforms the diffusion method in the one-port communication model; in particular, the ODE algorithm is found to be best suited for synchronous implementation in the static situation; and that the dimension exchange method is most superior in synchronous load balancing even under the all-port communication model; the strength of the diffusion method is in asynchronous implementation under the all-port communication model; the ODF algorithm performs best in high dimensional networks in this case. The rest of paper is organized as follows. Section 2 provides a framework of load balancing for our comparison of the various algorithms. Section 3 specifies load balancing algorithms in a unified forni. Section 4 compares load balancing algorithms when lhey are implemented in asynchronous and synchronous invocation policies. Section 5 reports the results from simulations that further assess the load balancing algorithms. We conclude in Section 6 with a summary of comparative results for the DE and the DF methods. 2 A generic model of load balancing The parallel computer we consider is composed of a finite set of homogeneous processors, which are interconnected by a direct communication network Processors communicate through passing messages. The communication channels are assumed to be full duplex so that a pair of directly connected (nearest neighbor ) processors can send/receive messages simultaneously t, )/from each other. In addition, we assume the sending and the receiving operation of a message in two ends of a channel take place instantaneously. We represent such a system by a simple connected graph G = (V, E), where V is a set of processors labeled 1 through N, and E' V x V is a set of edges. Every edge (i,j) E E corresponds to the communication channel between processors i and j. Let d(i) denote the set of nearest neighbors of processor i, d(i) = Id(;)/ be the degree of processor i, and d(g) be the maximum of d(i) for 15 i 5 N. The underlying parallel computation is assumed to comprise a large number of independent processes, which are the basic units of workload. The total number of processes are assumed to be large enough so that the workload of a processor is infinitely ditisible. Processes may be dynamically generated or consumed as the computation proceeds, and may also l)e migrated 473
4 across processors for the purpose of balancing. Correspondingly, we distinguish between two fundamental operations in a processor by their purposes: the computational operation and the balancing operation. An any time, a processor is performing a computational operation and/or balancing operation. Notice that during the execution of a balancing operation, the underlying computation can be suspended or performed concurrently. The concurrent execution of these two operations is possible when processors are capable of multiprogramming or the balancing operation is done in the background by cheap coprocessors. Since the workload of processors is either fixed or varying with time in the load balancing process, we refer to these two execution cases as SlQtZC arid dynamzc situations, respectively. Let t be an integer time variable, which is proportional to global real time. We quantify the workload of processor i at time t by w: in terms of the number of residing processes. Let z(t) denote the set of processors that are performing balancing operations at time t. Then, the change of workload of a processor at time t in the dynamic situation is modeled by the equation where 4fti denotes i,he aniounts of workload generated or finished from time t to t + 1, aiid ft(.) represents a load balancing operator. aft' = 0 in the static situation. This model is generic because the. balancing operator fl(.) and the set of processors in load balancing at any time t, z(t), are left undefined. The operator ft(.) can bc any nearest neighbor load balancing algorithms including the diffusion and the dimension exchange methods, which will be sperified in thc next section. The set I(t) is determined by invocation policies of load balancing. They are orthogonal tc) load balancing algorithms in that any invocation policy can he iniplemented in combination with any bad balancing algorit hm. Since a load balancing operat,ion incurs nonnegligible overheads, different applications require tliffcbrent invocation policies for better tradeoff between their benefits and extra overheads In parallel computations using domain decomposition techniques for example, the computational requirement. associated with each portion of a problem domain may didnge as the coiriputation proceeds. To reduce the pialty of load inibalarices, an effective way is to periodically redecoinpose the problem domain with the aini of achieving a. global uniform distribution arross processors. To this end, all processors are required to perform load balaiicing oyterations synchronously for a shtrrt period. That is, I(t) = { 1,2,,..., N } for t 3 to, where to is the instant when the whole system state satisfies certain conditions as those set in [9]. By contrast, the parallel execution of dynamic tree-structured computations usually requires only local balancing which assures no idle processors exist while there are other busy processors. Thus, each processor is allowed to invoke a load balancing operation at any time in an asynchronous manner according to its own local workload distribution. A simple policy is that once a processor's workload drops below a preset threshold, wunde+,ad, a load balancing operation is then activated. That, is, T(t) = {ilwj < Wunderload}. More sophisticated invocation policies were discussed in [lo, 21. Figure 1 presents an illustration of these two implementation models in a system of five processors. The dots and the triangles represent the computation operation and balancing operation, respectively. 3 The dimension exchange versus the diffusion methods This section briefly describe of the dimension exchange and the diffusion methods. Both of them are parameterized algorithms. We present several instances of these two methods based on different choices of values for their parameters. 3.1 The dimension exchange method With the dimension exchange method, any processor which invokes a load balancing operation balances its workload with its neighbors successively. I'or a processor i, it works in the following way that where j, E d(i), and 0 < X < 1, called the dimension exchange parameter, is preset to determine the fraction of excess workload to be migrated between a. pair of processors. The formula tells that a balancing operation in the dimension exchange method comprises d(i) pairwise balancing steps for processor i. At each step, processor i balances its workload with one of its neighbors, and uses the new result for the subsequent balancing. It is because of the sequential nature in the sequence of balancing steps, a load balancing operation requires d(i) communication steps in both the all-port and the oneport communication models. The efficiency of the DE method is determined by the dimension exchange parameter. A DE operation with different choices of the parameter will reduce the 474
5 c L % si E 2 a 2 I,,,,,,,,, I t t+5 t+10 t+ls time t t+5 t+10 t+15 time (a) Asynchronous implementation (b) Synchronous implementation Figure 1: An illustration of generic models of load balancing imbalance factor of a system state ty different factors. In t,he following, we present two choices of the parameter which have been suggested as rat,ional choices in the litcrature I Average damenbzon erthange (ADE) equally splits the total workload of a pair 01 processors by the choice X = 1/2 It is a straightforward choice for local balancing at each pairwise operation, and has been favored in hypercuhe netucrrks [3] 2 Optamally-funned dinif nsaon exchange (ODE) is a new variant of the DE method, which takes certain specific parameter values that have the effect of maximizing efficiency in global balancing [ll]. The optimal parameter depends on the topology and the size of underlying coinmunic ation network Let k = max{k,, 1 5 i 5 n} in the k~ x x C, mesh and torus 'Then, their optimal parameter values were shown, in [ll], as 0 X = 1/( 1 + sin(ir/k)) in the mesh, 0 X = I/( 1 + hin/h/k)) in the toriis 3.2 The diffusion method With the diffusion mrbthod, any processor which invokes a load balancing operation compares its workload with those of its nearest neighbors, and then gives away or takes in certain airioiint of workload with respect, to rac h nearest neighbor. The diffusion operator in a processor i can be written in the form that workload Iwi - wj I to processor j if wi > wj, or fetches some workload from processor j otherwise. Clearly, a load balancing operation with the I>F method requires only one communication step in t,he all-port, communication model, but d(i) steps in the one-port. communication model. As in the DE method, the efficiency of the DF method is determined by the diffusion parameter. Following are two common choices of the parameter. 1. Local average dzflusion (ADF) takes an average of workload of neighboring processors by setting a" zj - [12, 131. The torus is regular in t,hat processors have the same degree. The mesh is approximately regular when it is in large size. For simpiicit.y, we nse a single value cr = & to cover all communication channels in the mesh and the torus. 2. Opiimally-tuned dzflusion (ODF) is ii new variant of the DF method, which takes certain specific parameter values for maximizing efficiency in global balancing [3]. As in the DE method, the optimal diffusion parameter depends on the topology and the size of underlying networks. Let k = max{ IC1, k2,..., k,} in the IC1 x kz x I kn mesh and torus. Then, their optimal choices were shown, in [3, 141, as 0 LY = 1/2n in the mesh, 0 CY = l/(2n cos(27r/k)) in thc torus, 0 a = l/(n + 1) in the n-d hypercube. where 0 < cyzj < 1, called the diffusion parameter, is prt,defined to dictate the portion to be migrated bet,wchen any two processors. I'rocessor i apportions excess Assume t = 0 when processors invoke a synchronous or asynchronous load balancing procedure. We are concerned with subsequent workload distributions resulting 475
6 from different load balancing algorithms. Denote the overall workload distribution at certain time t by a vector Wt = (tu;, wi,..,u$,,). Denote its corresponding uniform distribution by a vector pt = (E', z',. ', G'), where E' = CL, wi/iv. We define a concept of system ambulance factor, denoted by ut, as the deviation of Wt from W'. That is, vt = ~ [Wt-~t~~' = E? a=] ( ~ 1 U f - - ~ w 1. The system imbalance factor reflects the variance of processors' workloads at a given point in time. With the system imbalance factor utl we define the efficiency of a load balancing algorithm, denoted by T, a5 the number of load balancing steps required to reduce the imbalance factor of the initial state to a tolerable level in the static situation; and define the balance quality as the bound for imbalancr factors which is to br guaranteed by the load balancing procedure III the dj namic situation. Load balancing algorithms will be compared with each other in terms of these two rneastires under the following assumptions. Throughout the section, E[.] denotes the expected value of a random variable. Assumption 4.1 At initial time, processors' workloads wp, 1 5 i 5 N, are N andepen,dent and identically distributed 6.i.d.) random variables with expectation,uo and variance U:. At any timet, t 2 0, processors' workload generation/'con~surription rataos &, 1 5 i < N, are zero in the static situation or i.i.d. random variables wrth expectation,u and variance U' zn the dynamic situ tit ion. 4.1 Asynchronous implementations In an asynchronous implementation of load balancing, processors perform load balancing operations discretely based on their own local workload distributions and invocation policies. Since load balancing algorithms can be treated as orthogonal to their invocation policies, we consider the load balancing operations of processors in one time step so as to isolate their effects on the sjstem imbalance factor from the effects of invocation policies. We focus OIL the static situation of load balancing in which the underlying computation in a processor i:, suspended while the processor is performing load balaiicing operations. The dynamic situation makes only a few differences to the analysis of the effects of load halancing. Let uo be the original system imbalance factor when t = 0, and U' be the system imbalance factor when t = 1. Our comparison will be made between vider uidf, and uidj which are resulting from various load balancing Operations Theorem 4.1 Suppose processors are running an asynchronous load balancing process under Assumption 4.1. Then, E[uAde] 5 E[ufif] in the one-port communication model, while E[uLtf] 5 E[uide] En the all-port communzcation model. Moreover, E[uidf] 5 E[uidf] in the chain and the rang networks but E[u,',,] 5 E[v&] in two- or higher- dimensional meshes and tori. In addition, E[uide] < E[u,lde] 2n the all-port cornmunicataon model. The comparison is based on a lemma concerning the sample variance of a combination of random variables in a sample set. We present it without proof. It can be easily shown using fundamental statistical theories. Lemma 4.1 Suppose that (1 (2,...,(N random variables with vanance U', Then, 1. for any k, 15 k < N, are N i.i.d and < = CE,(i. where 0 < a; < 1 satisfying ai = 1; and the variance is minimized at ai = l /k for a given k. 2. for any kl and k2 and 1 < k1 5 k2 5 N, where 0 < ai < 1 satisfying bi < 1 satisfying bj = 1. a; = 1 and 0 < Proof sketch of Theorem 4.1 At certain time in an asynchronous load balancing process, there might be more than one processor which are invoking load balajcing within their neighborhoods simultaneously. Let d(i) = {i} U d(i) denote the balancing domain of an invoker processor i. The balancing domains of concurrent invokers may be overlapping or separated with each other. As a whole, those processors which are running load balancing processes are partitioned into a number of separated spheres, some of which are singular balancing domains and some of which are unions of overlapping domains. Processors in different spheres perform load balancing operations independently, while processors in the same sphere perform load balancing in a synchronous manner. Suppose initially there are m independent balancing spheres in the system, denoted by B1, B2,..., Bm. 476
7 Then, by the definition of the system imbalance factor v, we have N The last term is a constant for a given number of prowssors in load balancing and independent of the topological relationships among the processors in load balancing. The first term is due to load balancing operations in all separated balancing spheres. It is a simple arithmetic sum of imbalance factors of each sphere, ~ ~ : r EE(lwf B J -- GII')). As a whole, E[v'] implies that the expected value of the system imbalance factor is influenced independently by load balancing operations within different balancing spheres Therefore, it suffices to compare the effects of load balancing algorithms within different spheres using Lemma 4.1 Owing to the limitation of space thc remainder of the proof is omitted here. This theorern says that the dimmsion exchange and the diffusion methods are suitable for the one-port and the all-port communication models, respectively. More specifically, it reveals that the ODF algorithm outperforms the ADE' algorithm in higher dimensional meshes arid tori although the ODF was originally proposed for iise in synchronous global balancing 4.2 Synchronous implementations In a synchronous implementation of load balancing, processors perform load halancing operations concurrently and continuously for a timc: period in order to achieve a global balanced state in the state situation or t(j keep the varying system imbalance factor bounded in the dynamic situation. From Eq. (2) and (3), it is kiiown that both the balancing operators, ti(.), of the DE and the DF methods are linear iterative operators. Hence, the synchronous implementation of Eq (1) can t l t b modeled by the equation N Wt+' = FWt (6) where F is either a IIE or a DF matrix defined by Eq. (2) or (3), respectively The features of synchronous implementations of the DE or DF methods are therefore fully captured by the iterative process governed by F. In the static situation, cpt = 0. According to fundamental iterative theoritas, we then have?'= 0(1/Iny(F)), (7) where y(f) is the subdominant eigenvalue of F in modulus. The closed expressions of y(f) are readily available in [12, 111 when the DE and the DF methods are applied in the mesh and the torus networks. Substituting them in Eq.(7), we obtain the efficiencies of the DE and the DF methods in both one-port and all-port communication models, as presented in Table 1. The entries of the table show that both the ADE and the ODE algorithms converge asymptotically faster than the diffusion method in the one-port communication model; and that in the all-port communication model, the ODE algorithm converges also faster than other three algorithms by a factor of k. In the dynamic situation, Eq.(7) leads to that E[d] = E(IIWt - Wf112) - E((pWt-1 - Vy) + E (pt- (a"12) = E(((Ff+lW" - t + CE(((p@+1-i - $+I- From Lemma 4.1, we then obtain that with the DE method in the one-port model, where b=(l-x)2+x2 ands=1+b+b2+...+bd-l; and with the DF method in the all-port model, 1 -- at+l E[4] = (U"".; + - I U i=o a2)n - (t + l)a2 - U;, where U = (1 - da)' + do2. Easily, we come to the following theorem. Theorem 4.2 Suppose processors are running synchronous DE and DF load balancing processes under Assumption 4.1. Then, E[.:&] 5 E[Y:~~], E[v:~~] 5 E[v:dj], and E[&] 5 E[&j] in both one-port and ailport communication models. 5 Experimental results In the preceding section, we explored a number of relationships between the dimension exchange and the diffusion methods with respect to their efficiencies and balancing qualities. In order to obtain an idea of the magnitude of their differencies, we conducted a statistical simulation of these load balancing algorithms on various topologies and sizes of communication networks and on synthetic workload distributions. The experimental results also serve to verify the theoretical results. 477
8 Table 1: Efficiencies of load balancing algorithms in the mesh and torus networks, where k is the maximum number of nodes over all dimensions in an n-d network and * - port means the all-port communication model torus mesh ADE ODE ADF ODF I 1--port8 *-port 1-port *-port 1-port *-port 1-port *-port O(nk2) O(nk2) O(nk) O(nk) O(n2k2) O(nkz) 0(n2k2) O(nk2) O(nk2) O(nk2) O(nk) O(nk) O(n2k2) O(nk2) O(n2k2) O(nk2) In the simulation, the initial workload distribution W is assumed to be a random vector, each element w of which is drawn independently from an idmtical uniform distribution in [0, looo] Each data point obtained in the experiment is the average of 20 runs, using dilfermi random initial workload distributions and different workload generation ratios We also assume thr underlying system is in all port communication model so that a DE balancing operation takes the time of 2n DF opt rations in the n-1) niesh and the n-d torus. A DF optaration is taken as a bmic time step in a load balanc ing process. The first experiment IS a simulittton of ss nchronous load balancing in the static behailor of workloads. In thc simulation, we measure the riuntber of cornrniinication steps, denoted by Z, necess,try for arriving at a tlobal balanced state We define the global balance stde as the state in which system imbalance factor is less than or equal to one Figure 2 plots the experiinental results from different load balancing algorithms in thfl 2-D mesh of various sims from 2 Y 2 to 32 x 32 change method out performs the diffusion ntethod even in the all-port communication model. In ptuticular, it is seen that the ODE algorithm accelerates the DE load balancing process significantly. In Figure 2, we also see that the number of communicatioii steps r in a 2-D niesh is dependent only on the size of it,s 1ii.rge dimension and insensitive to the size of its small dimension. This observation was proved to be true in both the mesh and the torus in [l 11. The second experiment is a simulation of asynchronous load balancing in the dynamic situation of random workload gent:rations/consumptions. In the simulation, we assume the expected workload generation ratio of a processor at each time step is 100 with the variance of 30 and the consumption ratio is a constant LOO. In the simulat,ion of asynchronous load balancing, we use a simple invocation policy that once a processor s workload drops or rises beyond a pair of preset bounds, LOO and 800, the processor then activates a load balancing operation. Figure 3 plot,s the system iml)alance factors resulting from different load balancing algorithms in a mesh of size 16 x I50 0 Figure 2: The number of necessary communication steps during a statically synchronous load balancing process in the 2-D mesh of various sizes from 2 x 2 to 32 x 32 This figure clearly indicates that t.he dimension ex- Figure 3: Change of the system imbalance factor in the first 200 steps of a dynamically asynchronous load balancing process in the rnesh of size 16 x 16 From this figure, it is seen that the ADE algorithm re- 478
9 duces the initial system imbalance factor more rapidly than the diffusion method and keeps it bounded in a much lower level. It can also be observed that both the ODE and the ODF algorithms, the optimally tuned algorithms for global synchronous load balancing, do not gain significant benefits in asynchronous implementations. 6 Conclusions In this paper, we made a comparison between two classes of nearest neighbor load balancing algorithms, tht. dimension exchange (DE) and the diffusion (DF) mcthods, with respect to their efficiency in driving any initial workload distribution to a uniform distribill ion and their ability in controlling the growth of variance among processors' workloads. We focused on thvir four instances -the ADE, the ODE. the ADF ant1 the ODF--which are the most common versions in practice. The comparison was made comprehensively in both one- port and all- 11 or t com mimic at ion models with consideration of various implementation strategieli: synchronous/asynchrorious invocation policies and stnticldynamic random workload behaviors. We showed that the DE method outperforms the DF method in rhe one-port, communication model. In particular, the ODE algorithm is best suited for synchronous implementation in the static situation. We also revealed of the superiority of the DE method in synchronous load balancing even in the all-port conimunication model. The strength of the diffusion method is in asynchronous implementation in the all-port cominunication model. The OIIF algorithm performs hest in high dimensional networks in that case. 'I'he comparative study not only provides an insight into nearest neighbor load balancing algorithms, but also offers practical guidelines to system developers in designing load balancing a1 chitectures for various parallc 1 computational paradigms. References [kj V. Kumar, A. Y. Grama, and N R. Vempaty. Scalable load balancing techniques for parallel computers. Journal of Parallel and Dastrzbuted Computmg, 22(1):60-79, July [2\ M. Willebeek-LeMair and A. P. Reeves. Strategies for dynamic load balancing on highly parallel computers. IEEE Transactzons on Parallel and Dastrzbuted Systems, 4(9):! , September [3] G. Cybenko. Load balancing for distributed memory multiprocessors. Journal of Parallel and Distributed Computing, 7: , [4] D. P. Bertsekas and J. N. Tsitsiklis. Parallel and distributed computation: Numerical methods, Prentice-Hall Inc., [5] B. Ghosh and S. Muthukrishnan. Dynamic load balancing in distributed networks by random matchings. In Proceedings of 6th ACM Symposium on Parallel Algorithms and Architectures, [6] C.-Z. Xu and F.C.M. Lau. Analysis of the generalized dimension exchange method for dynamic load balancing. Journal of Parallel and Distributed Computing, 16(4): , December [7] S. L. Johnsson and C.-T. Ho. Spanning graphs for optimum broadcasting and personalized communication in hypercubes. IEEE Transactions on Comput ers, 38( 9) : , Septembt :r [8] L. M. Ni and P. K. McKinley. A survey of wormhole routing techniques in direct networks. IEEE Computer, 26:62-76, February [9] D. M. Nicol and P. F. Reynolds. Optimal dynamic remapping of data parallel computation. IEEE Transactions on Computers, 39(2): , February [lo] R. Liiling and B. Monien. A dynamic distributed load balancing algorithm with provable good performance. In Proceedings of 5th ACM Symposium on Parallel Algorithms and Architectures, pages , [ll] C.-Z. Xu and F M. Lau. The generalized dimension exchange method for load balancing in %-ary n-cubes and variants. Journal of Pa.rallel and Distributed Computing, 21( 1):72-85, January [I21 J. B. Boillat. Load balancing and poisscm equation in a graph. Concurrency: Practice and Experience, 2(4): , December [13] X.3. Qian and Q. Yang. Load balancing on generalized hypercube and mesh multiprocessors with lal. In Proceedings of 11th International Conference on Distribuied Comp,uting Systems, pages , [14] C.-Z. Xu and F.C.M. Lau. Optimal parameters for load balancing with the diffusion method in mesh networks. Parallel Processing Letlers, 4(2): , June
Load Balancing Between Heterogenous Computing Clusters
Load Balancing Between Heterogenous Computing Clusters Siu-Cheung Chau Dept. of Physics and Computing, Wilfrid Laurier University, Waterloo, Ontario, Canada, N2L 3C5 e-mail: schau@wlu.ca Ada Wai-Chee Fu
More informationPerformance Comparison of Dynamic Load-Balancing Strategies for Distributed Computing
Performance Comparison of Dynamic Load-Balancing Strategies for Distributed Computing A. Cortés, A. Ripoll, M.A. Senar and E. Luque Computer Architecture and Operating Systems Group Universitat Autònoma
More informationLoad Balancing between Computing Clusters
Load Balancing between Computing Clusters Siu-Cheung Chau Dept. of Physics and Computing, Wilfrid Laurier University, Waterloo, Ontario, Canada, NL 3C5 e-mail: schau@wlu.ca Ada Wai-Chee Fu Dept. of Computer
More informationDECENTRALIZED LOAD BALANCING IN HETEROGENEOUS SYSTEMS USING DIFFUSION APPROACH
DECENTRALIZED LOAD BALANCING IN HETEROGENEOUS SYSTEMS USING DIFFUSION APPROACH P.Neelakantan Department of Computer Science & Engineering, SVCET, Chittoor pneelakantan@rediffmail.com ABSTRACT The grid
More informationExperiments on the local load balancing algorithms; part 1
Experiments on the local load balancing algorithms; part 1 Ştefan Măruşter Institute e-austria Timisoara West University of Timişoara, Romania maruster@info.uvt.ro Abstract. In this paper the influence
More informationThe Liquid Model Load Balancing Method
P APER ACCEPTED FOR THE JOURNAL OF P ARALLEL ALGORITHMS AND APPLICATIONS, SPECIAL ISSUE ON ALGORITHMS FOR ENHANCED MESH ARCHITECTURES The Liquid Model Load Balancing Method Dominik Henrich Institute for
More informationCSE 4351/5351 Notes 7: Task Scheduling & Load Balancing
CSE / Notes : Task Scheduling & Load Balancing Task Scheduling A task is a (sequential) activity that uses a set of inputs to produce a set of outputs. A task (precedence) graph is an acyclic, directed
More informationDistributed Dynamic Load Balancing for Iterative-Stencil Applications
Distributed Dynamic Load Balancing for Iterative-Stencil Applications G. Dethier 1, P. Marchot 2 and P.A. de Marneffe 1 1 EECS Department, University of Liege, Belgium 2 Chemical Engineering Department,
More informationTopological Properties
Advanced Computer Architecture Topological Properties Routing Distance: Number of links on route Node degree: Number of channels per node Network diameter: Longest minimum routing distance between any
More informationPerformance of Dynamic Load Balancing Algorithms for Unstructured Mesh Calculations
Performance of Dynamic Load Balancing Algorithms for Unstructured Mesh Calculations Roy D. Williams, 1990 Presented by Chris Eldred Outline Summary Finite Element Solver Load Balancing Results Types Conclusions
More informationLoad balancing in a heterogeneous computer system by self-organizing Kohonen network
Bull. Nov. Comp. Center, Comp. Science, 25 (2006), 69 74 c 2006 NCC Publisher Load balancing in a heterogeneous computer system by self-organizing Kohonen network Mikhail S. Tarkov, Yakov S. Bezrukov Abstract.
More informationEfficient Load Balancing by Adaptive Bypasses for the Migration on the Internet
Efficient Load Balancing by Adaptive Bypasses for the Migration on the Internet Yukio Hayashi yhayashi@jaist.ac.jp Japan Advanced Institute of Science and Technology ICCS 03 Workshop on Grid Computing,
More informationEnergy Efficient Load Balancing among Heterogeneous Nodes of Wireless Sensor Network
Energy Efficient Load Balancing among Heterogeneous Nodes of Wireless Sensor Network Chandrakant N Bangalore, India nadhachandra@gmail.com Abstract Energy efficient load balancing in a Wireless Sensor
More informationSystem Interconnect Architectures. Goals and Analysis. Network Properties and Routing. Terminology - 2. Terminology - 1
System Interconnect Architectures CSCI 8150 Advanced Computer Architecture Hwang, Chapter 2 Program and Network Properties 2.4 System Interconnect Architectures Direct networks for static connections Indirect
More informationFeedback guided load balancing in a distributed memory environment
Feedback guided load balancing in a distributed memory environment Constantinos Christofi August 18, 2011 MSc in High Performance Computing The University of Edinburgh Year of Presentation: 2011 Abstract
More informationA Network Flow Approach in Cloud Computing
1 A Network Flow Approach in Cloud Computing Soheil Feizi, Amy Zhang, Muriel Médard RLE at MIT Abstract In this paper, by using network flow principles, we propose algorithms to address various challenges
More informationCoupling Dynamic Load Balancing with Asynchronism in Iterative Algorithms on the Computational Grid
Coupling Dynamic Load Balancing with Asynchronism in Iterative Algorithms on the Computational Grid Jacques M. Bahi, Sylvain Contassot-Vivier and Raphaël Couturier Laboratoire d Informatique de Franche-Comté(LIFC),
More informationAn Empirical Study and Analysis of the Dynamic Load Balancing Techniques Used in Parallel Computing Systems
An Empirical Study and Analysis of the Dynamic Load Balancing Techniques Used in Parallel Computing Systems Ardhendu Mandal and Subhas Chandra Pal Department of Computer Science and Application, University
More informationKeywords: Dynamic Load Balancing, Process Migration, Load Indices, Threshold Level, Response Time, Process Age.
Volume 3, Issue 10, October 2013 ISSN: 2277 128X International Journal of Advanced Research in Computer Science and Software Engineering Research Paper Available online at: www.ijarcsse.com Load Measurement
More informationA Review of Customized Dynamic Load Balancing for a Network of Workstations
A Review of Customized Dynamic Load Balancing for a Network of Workstations Taken from work done by: Mohammed Javeed Zaki, Wei Li, Srinivasan Parthasarathy Computer Science Department, University of Rochester
More informationMapReduce and Distributed Data Analysis. Sergei Vassilvitskii Google Research
MapReduce and Distributed Data Analysis Google Research 1 Dealing With Massive Data 2 2 Dealing With Massive Data Polynomial Memory Sublinear RAM Sketches External Memory Property Testing 3 3 Dealing With
More informationLoad Balancing on Massively Parallel Networks: A Cellular Automaton Approach
Load Balancing on Massively Parallel Networks: A Cellular Automaton Approach by Brian Dickens at the Thomas Jefferson High School for Science and Technology supervising instructor: Donald Hyatt originally
More informationA Comparison of General Approaches to Multiprocessor Scheduling
A Comparison of General Approaches to Multiprocessor Scheduling Jing-Chiou Liou AT&T Laboratories Middletown, NJ 0778, USA jing@jolt.mt.att.com Michael A. Palis Department of Computer Science Rutgers University
More informationA Robust Dynamic Load-balancing Scheme for Data Parallel Application on Message Passing Architecture
A Robust Dynamic Load-balancing Scheme for Data Parallel Application on Message Passing Architecture Yangsuk Kee Department of Computer Engineering Seoul National University Seoul, 151-742, Korea Soonhoi
More informationHow To Balance In A Distributed System
6 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 11, NO. 1, JANUARY 2000 How Useful Is Old Information? Michael Mitzenmacher AbstractÐWe consider the problem of load balancing in dynamic distributed
More informationSHARP BOUNDS FOR THE SUM OF THE SQUARES OF THE DEGREES OF A GRAPH
31 Kragujevac J. Math. 25 (2003) 31 49. SHARP BOUNDS FOR THE SUM OF THE SQUARES OF THE DEGREES OF A GRAPH Kinkar Ch. Das Department of Mathematics, Indian Institute of Technology, Kharagpur 721302, W.B.,
More informationParallel Scalable Algorithms- Performance Parameters
www.bsc.es Parallel Scalable Algorithms- Performance Parameters Vassil Alexandrov, ICREA - Barcelona Supercomputing Center, Spain Overview Sources of Overhead in Parallel Programs Performance Metrics for
More informationAn Empirical Study of Two MIS Algorithms
An Empirical Study of Two MIS Algorithms Email: Tushar Bisht and Kishore Kothapalli International Institute of Information Technology, Hyderabad Hyderabad, Andhra Pradesh, India 32. tushar.bisht@research.iiit.ac.in,
More informationScheduling Allowance Adaptability in Load Balancing technique for Distributed Systems
Scheduling Allowance Adaptability in Load Balancing technique for Distributed Systems G.Rajina #1, P.Nagaraju #2 #1 M.Tech, Computer Science Engineering, TallaPadmavathi Engineering College, Warangal,
More informationAnalysis of an Artificial Hormone System (Extended abstract)
c 2013. This is the author s version of the work. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purpose or for creating
More informationDynamic Load Balancing in a Network of Workstations
Dynamic Load Balancing in a Network of Workstations 95.515F Research Report By: Shahzad Malik (219762) November 29, 2000 Table of Contents 1 Introduction 3 2 Load Balancing 4 2.1 Static Load Balancing
More informationA Hierarchical Structure based Coverage Repair in Wireless Sensor Networks
A Hierarchical Structure based Coverage Repair in Wireless Sensor Networks Jie Wu Computer Science & Engineering Department Florida Atlantic University Boca Raton, FL 3343, USA E-mail: jie@cse.fau.edu
More informationA Simultaneous Solution for General Linear Equations on a Ring or Hierarchical Cluster
Acta Technica Jaurinensis Vol. 3. No. 1. 010 A Simultaneous Solution for General Linear Equations on a Ring or Hierarchical Cluster G. Molnárka, N. Varjasi Széchenyi István University Győr, Hungary, H-906
More informationDYNAMIC LOAD BALANCING SCHEME FOR ITERATIVE APPLICATIONS
Journal homepage: www.mjret.in DYNAMIC LOAD BALANCING SCHEME FOR ITERATIVE APPLICATIONS ISSN:2348-6953 Rahul S. Wankhade, Darshan M. Marathe, Girish P. Nikam, Milind R. Jawale Department of Computer Engineering,
More informationA Review of Dynamic Load Balancing in Distributed Virtual En quantities
6 Dynamic Load Balancing in Distributed Virtual Environments using Heat Diffusion YUNHUA DENG and RYNSON W. H. LAU, City University of Hong Kong Distributed virtual environments (DVEs) are attracting a
More informationOn the Interaction and Competition among Internet Service Providers
On the Interaction and Competition among Internet Service Providers Sam C.M. Lee John C.S. Lui + Abstract The current Internet architecture comprises of different privately owned Internet service providers
More informationA New Nature-inspired Algorithm for Load Balancing
A New Nature-inspired Algorithm for Load Balancing Xiang Feng East China University of Science and Technology Shanghai, China 200237 Email: xfeng{@ecusteducn, @cshkuhk} Francis CM Lau The University of
More informationThe Answer Is Blowing in the Wind: Analysis of Powering Internet Data Centers with Wind Energy
The Answer Is Blowing in the Wind: Analysis of Powering Internet Data Centers with Wind Energy Yan Gao Accenture Technology Labs Zheng Zeng Apple Inc. Xue Liu McGill University P. R. Kumar Texas A&M University
More informationDYNAMIC GRAPH ANALYSIS FOR LOAD BALANCING APPLICATIONS
DYNAMIC GRAPH ANALYSIS FOR LOAD BALANCING APPLICATIONS DYNAMIC GRAPH ANALYSIS FOR LOAD BALANCING APPLICATIONS by Belal Ahmad Ibraheem Nwiran Dr. Ali Shatnawi Thesis submitted in partial fulfillment of
More informationA STUDY OF TASK SCHEDULING IN MULTIPROCESSOR ENVIROMENT Ranjit Rajak 1, C.P.Katti 2, Nidhi Rajak 3
A STUDY OF TASK SCHEDULING IN MULTIPROCESSOR ENVIROMENT Ranjit Rajak 1, C.P.Katti, Nidhi Rajak 1 Department of Computer Science & Applications, Dr.H.S.Gour Central University, Sagar, India, ranjit.jnu@gmail.com
More informationNumerical Analysis Lecture Notes
Numerical Analysis Lecture Notes Peter J. Olver 5. Inner Products and Norms The norm of a vector is a measure of its size. Besides the familiar Euclidean norm based on the dot product, there are a number
More informationTHE FUNDAMENTAL THEOREM OF ARBITRAGE PRICING
THE FUNDAMENTAL THEOREM OF ARBITRAGE PRICING 1. Introduction The Black-Scholes theory, which is the main subject of this course and its sequel, is based on the Efficient Market Hypothesis, that arbitrages
More informationCMSC 858T: Randomized Algorithms Spring 2003 Handout 8: The Local Lemma
CMSC 858T: Randomized Algorithms Spring 2003 Handout 8: The Local Lemma Please Note: The references at the end are given for extra reading if you are interested in exploring these ideas further. You are
More informationIntroduction to Scheduling Theory
Introduction to Scheduling Theory Arnaud Legrand Laboratoire Informatique et Distribution IMAG CNRS, France arnaud.legrand@imag.fr November 8, 2004 1/ 26 Outline 1 Task graphs from outer space 2 Scheduling
More informationBANACH AND HILBERT SPACE REVIEW
BANACH AND HILBET SPACE EVIEW CHISTOPHE HEIL These notes will briefly review some basic concepts related to the theory of Banach and Hilbert spaces. We are not trying to give a complete development, but
More informationLoad Balancing and Switch Scheduling
EE384Y Project Final Report Load Balancing and Switch Scheduling Xiangheng Liu Department of Electrical Engineering Stanford University, Stanford CA 94305 Email: liuxh@systems.stanford.edu Abstract Load
More informationA Simple Feature Extraction Technique of a Pattern By Hopfield Network
A Simple Feature Extraction Technique of a Pattern By Hopfield Network A.Nag!, S. Biswas *, D. Sarkar *, P.P. Sarkar *, B. Gupta **! Academy of Technology, Hoogly - 722 *USIC, University of Kalyani, Kalyani
More informationAn Ecient Dynamic Load Balancing using the Dimension Exchange. Ju-wook Jang. of balancing load among processors, most of the realworld
An Ecient Dynamic Load Balancing using the Dimension Exchange Method for Balancing of Quantized Loads on Hypercube Multiprocessors * Hwakyung Rim Dept. of Computer Science Seoul Korea 11-74 ackyung@arqlab1.sogang.ac.kr
More informationRESEARCH PAPER International Journal of Recent Trends in Engineering, Vol 1, No. 1, May 2009
An Algorithm for Dynamic Load Balancing in Distributed Systems with Multiple Supporting Nodes by Exploiting the Interrupt Service Parveen Jain 1, Daya Gupta 2 1,2 Delhi College of Engineering, New Delhi,
More informationSix Strategies for Building High Performance SOA Applications
Six Strategies for Building High Performance SOA Applications Uwe Breitenbücher, Oliver Kopp, Frank Leymann, Michael Reiter, Dieter Roller, and Tobias Unger University of Stuttgart, Institute of Architecture
More informationContributions to Gang Scheduling
CHAPTER 7 Contributions to Gang Scheduling In this Chapter, we present two techniques to improve Gang Scheduling policies by adopting the ideas of this Thesis. The first one, Performance- Driven Gang Scheduling,
More informationSIMULATION OF LOAD BALANCING ALGORITHMS: A Comparative Study
SIMULATION OF LOAD BALANCING ALGORITHMS: A Comparative Study Milan E. Soklic Abstract This article introduces a new load balancing algorithm, called diffusive load balancing, and compares its performance
More informationInterconnection Network
Interconnection Network Recap: Generic Parallel Architecture A generic modern multiprocessor Network Mem Communication assist (CA) $ P Node: processor(s), memory system, plus communication assist Network
More informationCompetitive Analysis of On line Randomized Call Control in Cellular Networks
Competitive Analysis of On line Randomized Call Control in Cellular Networks Ioannis Caragiannis Christos Kaklamanis Evi Papaioannou Abstract In this paper we address an important communication issue arising
More informationPerformance of networks containing both MaxNet and SumNet links
Performance of networks containing both MaxNet and SumNet links Lachlan L. H. Andrew and Bartek P. Wydrowski Abstract Both MaxNet and SumNet are distributed congestion control architectures suitable for
More informationChange Management in Enterprise IT Systems: Process Modeling and Capacity-optimal Scheduling
Change Management in Enterprise IT Systems: Process Modeling and Capacity-optimal Scheduling Praveen K. Muthusamy, Koushik Kar, Sambit Sahu, Prashant Pradhan and Saswati Sarkar Rensselaer Polytechnic Institute
More informationCost Model: Work, Span and Parallelism. 1 The RAM model for sequential computation:
CSE341T 08/31/2015 Lecture 3 Cost Model: Work, Span and Parallelism In this lecture, we will look at how one analyze a parallel program written using Cilk Plus. When we analyze the cost of an algorithm
More informationHyper Node Torus: A New Interconnection Network for High Speed Packet Processors
2011 International Symposium on Computer Networks and Distributed Systems (CNDS), February 23-24, 2011 Hyper Node Torus: A New Interconnection Network for High Speed Packet Processors Atefeh Khosravi,
More informationParallel Programming
Parallel Programming Parallel Architectures Diego Fabregat-Traver and Prof. Paolo Bientinesi HPAC, RWTH Aachen fabregat@aices.rwth-aachen.de WS15/16 Parallel Architectures Acknowledgements Prof. Felix
More informationDATA ANALYSIS II. Matrix Algorithms
DATA ANALYSIS II Matrix Algorithms Similarity Matrix Given a dataset D = {x i }, i=1,..,n consisting of n points in R d, let A denote the n n symmetric similarity matrix between the points, given as where
More informationCompletion Time Scheduling and the WSRPT Algorithm
Completion Time Scheduling and the WSRPT Algorithm Bo Xiong, Christine Chung Department of Computer Science, Connecticut College, New London, CT {bxiong,cchung}@conncoll.edu Abstract. We consider the online
More informationThe Heat Equation. Lectures INF2320 p. 1/88
The Heat Equation Lectures INF232 p. 1/88 Lectures INF232 p. 2/88 The Heat Equation We study the heat equation: u t = u xx for x (,1), t >, (1) u(,t) = u(1,t) = for t >, (2) u(x,) = f(x) for x (,1), (3)
More informationAperiodic Task Scheduling
Aperiodic Task Scheduling Jian-Jia Chen (slides are based on Peter Marwedel) TU Dortmund, Informatik 12 Germany Springer, 2010 2014 年 11 月 19 日 These slides use Microsoft clip arts. Microsoft copyright
More informationTHREE DIMENSIONAL GEOMETRY
Chapter 8 THREE DIMENSIONAL GEOMETRY 8.1 Introduction In this chapter we present a vector algebra approach to three dimensional geometry. The aim is to present standard properties of lines and planes,
More informationAdaptive Search with Stochastic Acceptance Probabilities for Global Optimization
Adaptive Search with Stochastic Acceptance Probabilities for Global Optimization Archis Ghate a and Robert L. Smith b a Industrial Engineering, University of Washington, Box 352650, Seattle, Washington,
More informationLecture 4 Online and streaming algorithms for clustering
CSE 291: Geometric algorithms Spring 2013 Lecture 4 Online and streaming algorithms for clustering 4.1 On-line k-clustering To the extent that clustering takes place in the brain, it happens in an on-line
More informationOnline Scheduling with Bounded Migration
Online Scheduling with Bounded Migration Peter Sanders, Naveen Sivadasan, and Martin Skutella Max-Planck-Institut für Informatik, Saarbrücken, Germany, {sanders,ns,skutella}@mpi-sb.mpg.de Abstract. Consider
More informationFairness in Routing and Load Balancing
Fairness in Routing and Load Balancing Jon Kleinberg Yuval Rabani Éva Tardos Abstract We consider the issue of network routing subject to explicit fairness conditions. The optimization of fairness criteria
More informationSystolic Computing. Fundamentals
Systolic Computing Fundamentals Motivations for Systolic Processing PARALLEL ALGORITHMS WHICH MODEL OF COMPUTATION IS THE BETTER TO USE? HOW MUCH TIME WE EXPECT TO SAVE USING A PARALLEL ALGORITHM? HOW
More information1 Norms and Vector Spaces
008.10.07.01 1 Norms and Vector Spaces Suppose we have a complex vector space V. A norm is a function f : V R which satisfies (i) f(x) 0 for all x V (ii) f(x + y) f(x) + f(y) for all x,y V (iii) f(λx)
More informationHeat Diffusion Based Dynamic Load Balancing for Distributed Virtual Environments
Heat Diffusion Based Dynamic Load Balancing for Distributed Virtual Environments Yunhua Deng Rynson W.H. Lau Department of Computer Science, City University of Hong Kong, Hong Kong Abstract Distributed
More informationA Novel Switch Mechanism for Load Balancing in Public Cloud
International OPEN ACCESS Journal Of Modern Engineering Research (IJMER) A Novel Switch Mechanism for Load Balancing in Public Cloud Kalathoti Rambabu 1, M. Chandra Sekhar 2 1 M. Tech (CSE), MVR College
More informationOPTIMAL DESIGN OF DISTRIBUTED SENSOR NETWORKS FOR FIELD RECONSTRUCTION
OPTIMAL DESIGN OF DISTRIBUTED SENSOR NETWORKS FOR FIELD RECONSTRUCTION Sérgio Pequito, Stephen Kruzick, Soummya Kar, José M. F. Moura, A. Pedro Aguiar Department of Electrical and Computer Engineering
More informationSingle machine parallel batch scheduling with unbounded capacity
Workshop on Combinatorics and Graph Theory 21th, April, 2006 Nankai University Single machine parallel batch scheduling with unbounded capacity Yuan Jinjiang Department of mathematics, Zhengzhou University
More informationA Hybrid Load Balancing Policy underlying Cloud Computing Environment
A Hybrid Load Balancing Policy underlying Cloud Computing Environment S.C. WANG, S.C. TSENG, S.S. WANG*, K.Q. YAN* Chaoyang University of Technology 168, Jifeng E. Rd., Wufeng District, Taichung 41349
More informationEnergy Efficient MapReduce
Energy Efficient MapReduce Motivation: Energy consumption is an important aspect of datacenters efficiency, the total power consumption in the united states has doubled from 2000 to 2005, representing
More informationUSING SPECTRAL RADIUS RATIO FOR NODE DEGREE TO ANALYZE THE EVOLUTION OF SCALE- FREE NETWORKS AND SMALL-WORLD NETWORKS
USING SPECTRAL RADIUS RATIO FOR NODE DEGREE TO ANALYZE THE EVOLUTION OF SCALE- FREE NETWORKS AND SMALL-WORLD NETWORKS Natarajan Meghanathan Jackson State University, 1400 Lynch St, Jackson, MS, USA natarajan.meghanathan@jsums.edu
More informationDecentralized Utility-based Sensor Network Design
Decentralized Utility-based Sensor Network Design Narayanan Sadagopan and Bhaskar Krishnamachari University of Southern California, Los Angeles, CA 90089-0781, USA narayans@cs.usc.edu, bkrishna@usc.edu
More informationPriority Based Load Balancing in a Self-Interested P2P Network
Priority Based Load Balancing in a Self-Interested P2P Network Xuan ZHOU and Wolfgang NEJDL L3S Research Center Expo Plaza 1, 30539 Hanover, Germany {zhou,nejdl}@l3s.de Abstract. A fundamental issue in
More informationParallel Programming at the Exascale Era: A Case Study on Parallelizing Matrix Assembly For Unstructured Meshes
Parallel Programming at the Exascale Era: A Case Study on Parallelizing Matrix Assembly For Unstructured Meshes Eric Petit, Loïc Thebault, Quang V. Dinh May 2014 EXA2CT Consortium 2 WPs Organization Proto-Applications
More informationReliable Systolic Computing through Redundancy
Reliable Systolic Computing through Redundancy Kunio Okuda 1, Siang Wun Song 1, and Marcos Tatsuo Yamamoto 1 Universidade de São Paulo, Brazil, {kunio,song,mty}@ime.usp.br, http://www.ime.usp.br/ song/
More information! Solve problem to optimality. ! Solve problem in poly-time. ! Solve arbitrary instances of the problem. #-approximation algorithm.
Approximation Algorithms 11 Approximation Algorithms Q Suppose I need to solve an NP-hard problem What should I do? A Theory says you're unlikely to find a poly-time algorithm Must sacrifice one of three
More informationScalability and Classifications
Scalability and Classifications 1 Types of Parallel Computers MIMD and SIMD classifications shared and distributed memory multicomputers distributed shared memory computers 2 Network Topologies static
More informationAn Efficient Distributed Algorithm to Identify and Traceback DDoS Traffic
Ó The Author 26. Published by Oxford University Press on behalf of The British Computer Society. All rights reserved. For Permissions, please email: journals.permissions@oxfordjournals.org doi:1.193/comjnl/bxl26
More informationFOUNDATIONS OF ALGEBRAIC GEOMETRY CLASS 22
FOUNDATIONS OF ALGEBRAIC GEOMETRY CLASS 22 RAVI VAKIL CONTENTS 1. Discrete valuation rings: Dimension 1 Noetherian regular local rings 1 Last day, we discussed the Zariski tangent space, and saw that it
More informationCOUNTING INDEPENDENT SETS IN SOME CLASSES OF (ALMOST) REGULAR GRAPHS
COUNTING INDEPENDENT SETS IN SOME CLASSES OF (ALMOST) REGULAR GRAPHS Alexander Burstein Department of Mathematics Howard University Washington, DC 259, USA aburstein@howard.edu Sergey Kitaev Mathematics
More informationLABEL PROPAGATION ON GRAPHS. SEMI-SUPERVISED LEARNING. ----Changsheng Liu 10-30-2014
LABEL PROPAGATION ON GRAPHS. SEMI-SUPERVISED LEARNING ----Changsheng Liu 10-30-2014 Agenda Semi Supervised Learning Topics in Semi Supervised Learning Label Propagation Local and global consistency Graph
More informationThis article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
IEEE/ACM TRANSACTIONS ON NETWORKING 1 A Greedy Link Scheduler for Wireless Networks With Gaussian Multiple-Access and Broadcast Channels Arun Sridharan, Student Member, IEEE, C Emre Koksal, Member, IEEE,
More informationState of Stress at Point
State of Stress at Point Einstein Notation The basic idea of Einstein notation is that a covector and a vector can form a scalar: This is typically written as an explicit sum: According to this convention,
More informationOn the Traffic Capacity of Cellular Data Networks. 1 Introduction. T. Bonald 1,2, A. Proutière 1,2
On the Traffic Capacity of Cellular Data Networks T. Bonald 1,2, A. Proutière 1,2 1 France Telecom Division R&D, 38-40 rue du Général Leclerc, 92794 Issy-les-Moulineaux, France {thomas.bonald, alexandre.proutiere}@francetelecom.com
More informationSection 1.1. Introduction to R n
The Calculus of Functions of Several Variables Section. Introduction to R n Calculus is the study of functional relationships and how related quantities change with each other. In your first exposure to
More informationJan F. Prins. Work-efficient Techniques for the Parallel Execution of Sparse Grid-based Computations TR91-042
Work-efficient Techniques for the Parallel Execution of Sparse Grid-based Computations TR91-042 Jan F. Prins The University of North Carolina at Chapel Hill Department of Computer Science CB#3175, Sitterson
More informationData Management in Hierarchical Bus Networks
Data Management in Hierarchical Bus Networks F. Meyer auf der Heide Department of Mathematics and Computer Science and Heinz Nixdorf Institute, Paderborn University, Germany fmadh@upb.de H. Räcke Department
More informationPrinciples and characteristics of distributed systems and environments
Principles and characteristics of distributed systems and environments Definition of a distributed system Distributed system is a collection of independent computers that appears to its users as a single
More informationPreserving Message Integrity in Dynamic Process Migration
Preserving Message Integrity in Dynamic Process Migration E. Heymann, F. Tinetti, E. Luque Universidad Autónoma de Barcelona Departamento de Informática 8193 - Bellaterra, Barcelona, Spain e-mail: e.heymann@cc.uab.es
More informationHardware-Aware Analysis and. Presentation Date: Sep 15 th 2009 Chrissie C. Cui
Hardware-Aware Analysis and Optimization of Stable Fluids Presentation Date: Sep 15 th 2009 Chrissie C. Cui Outline Introduction Highlights Flop and Bandwidth Analysis Mehrstellen Schemes Advection Caching
More informationLoad Balancing in MapReduce Based on Scalable Cardinality Estimates
Load Balancing in MapReduce Based on Scalable Cardinality Estimates Benjamin Gufler 1, Nikolaus Augsten #, Angelika Reiser 3, Alfons Kemper 4 Technische Universität München Boltzmannstraße 3, 85748 Garching
More informationPart 2: Community Detection
Chapter 8: Graph Data Part 2: Community Detection Based on Leskovec, Rajaraman, Ullman 2014: Mining of Massive Datasets Big Data Management and Analytics Outline Community Detection - Social networks -
More informationDistributed Computing over Communication Networks: Topology. (with an excursion to P2P)
Distributed Computing over Communication Networks: Topology (with an excursion to P2P) Some administrative comments... There will be a Skript for this part of the lecture. (Same as slides, except for today...
More informationLoad Balancing In Concurrent Parallel Applications
Load Balancing In Concurrent Parallel Applications Jeff Figler Rochester Institute of Technology Computer Engineering Department Rochester, New York 14623 May 1999 Abstract A parallel concurrent application
More information